Developer open-sources Growth-Harness, an agent framework without a hard turn cap

On 2026-11-05, a developer posted about an open-source project called Growth-Harness on V2EX, describing it as a non-parametric self-evolution mechanism for AI agents. The project has been fully open-sourced, with a link provided to an accompanying WeChat article. The core of the framework is its runtime, which does not impose a hard round cap. The author explains the design decision as follows. Most agent harnesses attach a hard stop condition to tasks, and they typically hard-code that mechanism, which the author calls not very AI native. In complex scenarios, task boundaries are often fuzzy, so an agent needs to execute long chains of steps, manage its own context, recover from failures on its own, and resume when a task flow is interrupted rather than starting over from scratch. In Growth-Harness, the question of how long a task should run is left to the agent itself. Every 30 rounds, the agent pauses and audits its own progress, reviewing what has been completed and what remains, then decides whether it is worth continuing. If it is, the run goes on; if not, the run wraps up. The author also states that the real constraint is not the number of rounds but how many tokens you dare to give the AI. Alongside the runtime, the project proposes a non-parametric pre-registered learning self-evolution engine. The author plans to verify the project by running a continuous 30-day scenario on local compute once the next mac mini arrives. This V2EX post is the only public report on the project so far, and that verification has not yet begun.